Skip to content
  • Florian Westphal's avatar
    netlink: remove mmapped netlink support · d1b4c689
    Florian Westphal authored
    mmapped netlink has a number of unresolved issues:
    
    - TX zerocopy support had to be disabled more than a year ago via
      commit 4682a035 ("netlink: Always copy on mmap TX.")
      because the content of the mmapped area can change after netlink
      attribute validation but before message processing.
    
    - RX support was implemented mainly to speed up nfqueue dumping packet
      payload to userspace.  However, since commit ae08ce00
      ("netfilter: nfnetlink_queue: zero copy support") we avoid one copy
      with the socket-based interface too (via the skb_zerocopy helper).
    
    The other problem is that skbs attached to mmaped netlink socket
    behave different from normal skbs:
    
    - they don't have a shinfo area, so all functions that use skb_shinfo()
    (e.g. skb_clone) cannot be used.
    
    - reserving headroom prevents userspace from seeing the content as
    it expects message to start at skb->head.
    See for instance
    commit aa3a0220 ("netlink: not trim skb for mmaped socket when dump").
    
    - skbs handed e.g. to netlink_ack must have non-NULL skb->sk, else we
    crash because it needs the sk to check if a tx ring is attached.
    
    Also not obvious, leads to non-intuitive bug fixes such as 7c7bdf35
    ("netfilter: nfnetlink: use original skbuff when acking batches").
    
    mmaped netlink also didn't play nicely with the skb_zerocopy helper
    used by nfqueue and openvswitch.  Daniel Borkmann fixed this via
    commit 6bb0fef4
    
     ("netlink, mmap: fix edge-case leakages in nf queue
    zero-copy")' but at the cost of also needing to provide remaining
    length to the allocation function.
    
    nfqueue also has problems when used with mmaped rx netlink:
    - mmaped netlink doesn't allow use of nfqueue batch verdict messages.
      Problem is that in the mmap case, the allocation time also determines
      the ordering in which the frame will be seen by userspace (A
      allocating before B means that A is located in earlier ring slot,
      but this also means that B might get a lower sequence number then A
      since seqno is decided later.  To fix this we would need to extend the
      spinlocked region to also cover the allocation and message setup which
      isn't desirable.
    - nfqueue can now be configured to queue large (GSO) skbs to userspace.
      Queing GSO packets is faster than having to force a software segmentation
      in the kernel, so this is a desirable option.  However, with a mmap based
      ring one has to use 64kb per ring slot element, else mmap has to fall back
      to the socket path (NL_MMAP_STATUS_COPY) for all large packets.
    
    To use the mmap interface, userspace not only has to probe for mmap netlink
    support, it also has to implement a recv/socket receive path in order to
    handle messages that exceed the size of an rx ring element.
    
    Cc: Daniel Borkmann <daniel@iogearbox.net>
    Cc: Ken-ichirou MATSUZAWA <chamaken@gmail.com>
    Cc: Pablo Neira Ayuso <pablo@netfilter.org>
    Cc: Patrick McHardy <kaber@trash.net>
    Cc: Thomas Graf <tgraf@suug.ch>
    Signed-off-by: default avatarFlorian Westphal <fw@strlen.de>
    Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
    d1b4c689